The sensitivity of financial distress prediction models to departures from normality*

Abstract

Abstract. This research empirically investigated the effect of nonnormality on financial stress prediction. The analysis included the application of prohit, logit and multiple discriminant analysis to prediction models found in previous literature, and also involved separate samples for both bankrupt and prohlem-status companies. Finally, the statistical techniques were evaluated under extreme conditions of nonnormality.Two basic procedures were used to modify the ratio distdhutions to achieve normality. These included a square-root transformation procedure and an outlier deletion procedure. Results were compared using hoth a univariate and a multivariate technique to identify and remove outliers. The results indicate the general sensitivity of the multiple discriminant analysis technique to departures from normality and the sensitivity of the logit and probit tectuiiques to extreme nonnormality. The data indicate that researchers interested in assessing classification accuracy might benefit by testing for distrihutional sensitivity using procedures outlined in this researeh.Risutni. Les auteurs ont proc^d6 k une analyse empirique de l'incidence des 6carts par rapport h. la nonnalit6 sur la provision des contraintes financi&res. L'analyse comporte Tapplication du probit, du logit et de l'analyse ^ discriminants multiples aux modMes pr^visionnnels que Ton trouve dans des publications, et l'on a eu recours ^ des ^chantillons distincts tant pour les socidt^s en situation de faillite que pour les soci6t6s en difficultd. Enfin, Ies techniques statistiques ont ii& 6valu6es dans des conditions extremes d'6cart par rapport ^ la normality.Deux m6thodes fondamentales ont €vt utilisdes pour modifier les distributions de ratios de fagon k parvenir & la normality. Ces methodes compiennent un proc^d^ de transformation de la racine can^e et un proc6d6 d'61imination des 6Mments isolds. Les rdsultats obtenus ont 6t6 compares ^ I'aide d'une technique univari6e ainsi que d'une technique multivari6e pour rep6reret supprimer les 616ments isol^s. Les rfisultats indiquent la sensibiht6 g6n&ale de la technique d'analyse'^ discriminants multiples aux ddviations par rapport & la normality et la sensihilitd des techniques logit et probit aux hearts extremes par rapport k la normality. Ces donn6es rdvfelent que les chercheurs qui s'int^ressent a l'^v^uation de l'axactitude de la classification pourraient tirer profit d'une verification de la sensibility de la distribution au moyen des mdthodes ddcrites par les auteurs. Financial Distress Prediction Models 285Introduction A major stream of research in the accounting literature has focused on the prediction of financial distress using linear statistical models and financial ratios (eg. Altman, Haldeman and Narayanan (1977), Ohlson (1980), and summaries in Zavgren (1983) and Foster (1986 ch. 15)). Frecka and Hopwood (1983) observed that financial ratios are typically skewed and nonnonnally distributed and that, in some cases, the nonnormality can be very extreme. They suggested that the prob...

Inability of Gearing-Ratio as Predictor for Early Warning Systems

“…It seems that strong deviations from normality of data are influencing the estimation procedure of logistic regression and are affecting the classification accuracy of logistic functions. This result is consistent with some other prior studies (Hopwood et al, 1988;Silva et al, 2002). Therefore, it cannot be concluded that logistic regression is superior for model building in contrast to discriminant analysis, which is in congruence with several findings from previous research (Casey and Moreover, the unequal distribution between bankrupt and non-bankrupt firms seems to influence the classification results of both models.…”

“…Especially for discriminant analysis it seems to be relevant that normally distributed data is available, because this is a theoretical pre-condition for proper application of this method (Klecka, 1980, p. 61;Hopwood, McKeown and Mutchler, 1988;Subhash, 1996, p. 263). Nevertheless, several results provided evidence that a weak violation of normality assumptions is not affecting the prediction accuracy of the final model that much, so that some departures can be argued (Hopwood et al, 1988;Silva, Stam and Neter, 2002). In some cases departures are beneficial for better discrimination in means, which can lead to better classification results compared to logistic regression (Pohar, Blas and Turk, 2004).…”

The Relevance of Trend Variables for the Prediction of Corporate Crises and Insolvencies

“…This aspect is necessary, because one important pre-condition for the correct application of multivariate linear discriminant analysis is the availability of normally distributed data (Klecka, 1980, p. 61;Subhash, 1996, p. 263;Hopwood et al, 1988a). Nevertheless, a small deviation from normality can be accepted as this aspect does not infl uence the classifi cation accuracy of the forecasting models all that heavily (Hopwood et al, 1988a;Feldesman, 2002;Silva et al, 2002). Based on the results shown in the Figures 6 and 7 in the appendix, it can be seen that for almost all accounting ratios and trends, the assumption of normality cannot be presumed.…”

Audit Firm Tenure and Auditor Reporting Quality: Evidence in Malaysia

“…This is due to the fact that, the higher probability of bankruptcy, the higher the need of the auditors to issue going-concern opinion. Regardless of whatever bankruptcy model being employed in prior researches (Hopwood, McKeown & Mutchler, 1989;Vanstraelen, 2000) in going concern opinion, the results suggest that auditors do assess distress condition of their clients. Prior research in Malaysia by Md.…”